Blog

Unsupervised Machine Learning Models 

– Overview of unsupervised machine learning

• Clustering
• K means
• Spectral clustering
• Partitioning methods
• Hierarchical-based methods
• Principal Component Analysis
• Constraint-based methods
• Mean-shift clustering
• DBSCAN
• Agglomerative hierarchical clustering
• Association rules
• AIS and SETM
• Apriori
• Equivalence Class Clustering and bottom-up Lattice Traversal
• FP-Growth

Conclusive remark on machine learning application in stock market prices prediction

Unsupervised machine learning

Unsupervised machine learning mainly refers to the algorithms used in the identification of patterns within the data set entailing the data points considered to be neither labeled nor classified. Hence, the algorithms are allowed to label, classify, and group such data points in the data set without any external guidance in executing such tasks. Therefore, the users are not expected to supervise the models (Pahwa et al., 2017). This implies that no labels are assigned to the learning algorithms, leaving it to find the structure within the input set. Hence, in unsupervised learning, artificial intelligence would classify the unsorted stock data aligned to differences and similarities, although there are no groups offered in the input data (Sakhare & Imambi, 2019). The primary objective of unsupervised learning is discovering the interesting and hidden patterns within the unlabeled data set, which investors can use in making informed decisions regarding the stock market. Unsupervised learning is categorized in two that is clustering and association rule algorithms that can be applied in forecasting stock.

Clustering learning

The cluster analysis, clustering, is a machine learning technique used to classify and identify related data points within a huge dataset without concern for the particular result. It essentially groups the collection of stock data in that the information within a similar category, known as a cluster, is essentially more similar to other compared to other groups (Prasdika & Sugiantoro, 2018). Frequently, it is applied as the data analysis method of discovering interesting stock trends or patterns, such as a group of companies based on their stock value (Bini & Mathew, 2016). There are various clustering methods, including K means, Spectral clustering, and Partitioning methods.

K-Means

K Means clustering is an unsupervised machine learning method and extremely simple technique of classifying individual data points into clusters (Zhu & Liu, 2021). The main objective of the method is to group the points into clusters, whereby every group is linked to its center mass. However, its main challenge is the identification of the centers and group of the mass, K-centroids. Further, every data point is governed by distance from the cluster’s centers and allows the unclassified data to be classified into groups. Additionally, with the predetermined k, the K Means algorithm proceeds by alternating betwixt two phases, the update and assignment phases. The update phase utilizes the results of the assignment phase in calculating the novel means, centroids of the newly established clusters, whereas the assignment phase mainly assigns every example to its closest cluster (Sinaga & Yang, 2020). Thus, provided the set of the individual data points portrayed by vectors, in which every vector entry represents a feature. Further, when all the objects in the data have been clustered and assigned into groups, the position of the k centroids should be recalculated to become the new cluster centers for the next repetition. The iteration is repeated until it merges, and the centroids no longer move (Fabregas, Gerardo & Tanguilig III, 2017).

The algorithm is utilized in trading grounded on the trend forecast strategy, which entails three main steps: partitioning, analysis, and forecast (Chen & Liu, 2020). Further, the K-means cluster algorithm is utilized to partition the stock cost time sequence information, after which linear regression is applied in analyzing the trend in every cluster (Bini & Mathew, 2016). The outcome of the analysis is then applied in trend forecast for windowed time sequence information. After processing, the information is forwarded to a machine learning algorithm for model creation. Thus, in the K-means clustering, K is the user input that is applied to create the k number for the classes. The advantage of the K-means over other clustering algorithms is the user-defined cluster number (Fithri & Wardhana, 2021). For instance, clusters can be formed that are responsible for the purchase, sell, and hold decisions.

Spectral clustering algorithm

Spectral clustering is a developing cluster algorithm that has performed exceptionally compared to numerous traditional clustering algorithms in various situations. This technique treats every data point to represent a graph node, changing the cluster issue to produce the graph-partitioning issue. Therefore, the usual adaptation entails three basic phases: building the similarity graph, projecting the information onto the lower-dimensional space, and clustering the information (Huang et al., 2019). The initial step structures the similarity graph to produce the adjacency matrix typically represented by A. the graph is built using the K-Nearest Neighbors, Epsilon-neighborhood graph, and fully connected graph (Tremblay & Loukas, 2020). The second phase is executed to constitute all the probability that members within a similar cluster could be far away within a particular dimensional space. Therefore, the space can be decreased to ensure that such points are nearer within the abridged space and be calculated accumulatively by applying the traditional culturing algorithm (Zhu et al., 2018). This can be done by calculating the graph Laplacian matrix, which is then normalized for efficiency. Additionally, the eigenvalues and vectors are computed to ensure a decrease in dimension. If the clusters’ number is k, the initial vectors and values are combined to form a matrix (Kerenidis & Landman, 2021). The last phase involves clustering information by applying any traditional cluster method, K-means clustering. Initially, every node is assigned a row of the normalized graph Laplacian matrix. After which, the information is clustered utilizing any traditional method. To transform the clustering outcome, the node identifier is retained.

Partitioning method

The partitioning technique creates the K clusters or partitions with ‘x’ data objects from a particular dataset. This algorithm begins operating by mainly assigning various objects to distinct datasets arbitrarily, after which, at every repetition, it recollects the information objects to other partitions (Lu & Zhu, 2017). Further, every partition is shown as a centroid, which is the average of every the objects within the partition. Each partition should entail a single information object in this segment, and every data object should contain just a single cluster (Swarndeep Saket & Pandya, 2016). The most universal cluster algorithm founded on partitioning techniques includes K-means, CLARA, and K-Mediods (Cheng, 2017). Thus, the traders can use the K-means and K-Mediods to predict stock values.

Hierarchical clustering algorithm

The algorithm is an unsupervised machine learning method aimed at finding the natural grouping founded on the parameters of the data. Also, it finds nested clusters of the information by building the required hierarchy. This technique creates a hierarchical decomposition of a specific set of data objects. It can be categorized as divisive or agglomerative, founded on how the hierarchical decomposition is formulated. The agglomerative strategy is the bottom-up tactic that begins with every object forming a separate group. Additionally, the hierarchical algorithms create a hierarchical decomposition of the particular dataset objects. Besides, the hierarchical decomposition is portrayed by the tree structure known as a dendrogram and does not require clusters as its input. In this form of clustering, it is plausible to view the partitions at a distinct level of granularities utilizing various forms of K, like in flat clustering (Shetty & Singh, 2021). After which, it combines the close groups until all the groups are merged into one. On the other hand, a divisive strategy is a top-down strategy that begins with all the clusters within the same group. Then in every iteration phase, a cluster is split into smaller groups until every object is within one cluster (Roux, 2018). Additionally, this technique encompasses the CURE and BIRCH algorithms. However, the defect of this technique is that once the split or merge is attained, it cannot be undone or even exchanged between the clustering objects. Thus, it cannot possibly correct the wrong decisions made. The traders can then use parametric clustering such as K-Means to categorize the cluster into two sections. For every cluster, the traders can further divide it down into two groups until they attain the desired number of clusters (Shah, Isah & Zulkernine, 2019). The technique is perceived to be extremely effective, whereas non-determinism is an issue in flat cluster. Moreover, hierarchical agglomerative cluster is applied in the stock prediction, resulting in a more useful structure compared to flat clustering found on the unstructured information. Additionally, hierarchy offers insight to investors for effective management categories. Further, for a long-term tactical decision, the investor should consider both present and direct competition from similar entities and the indirect competitions, which come from the entity variants and could be a considerable threat. Besides, the hierarchy extracted by such techniques could be further deemed a decision tree, which could aid in classifying novel items (Sokol, 2019). Nevertheless, the technique needs high computational power, as well as memory utilization, as they are founded on the creation of high dimensional distance matrices utilized for pairwise comparisons betwixt all the availed data points.

Principal Component Analysis

The principal component analysis, PCA, is a technique utilized in minimizing the dimensions of various variables within the particular set of data. It mainly entails the application of the covariance analysis amongst various parameters. In the stock prediction, the original information provided is mapped into the novel coordinates system founded on the variance in the specific market data (Wen, Lin & Nie, 2020). Further, the PCA uses a statistical process to transmute correlated variables into linearly uncorrelated variables known as the principal components. Therefore, to decrease the dimensionality of transformed information, just minor components are deemed since the initial principal component is perceived to account for the vast variable within the information and every succeeding component accounts for most of the remaining inconsistency (Singh & Khushi, 2021). This technique is essential when there is a large correlated dimension, which always entails data redundancy. Therefore, PCA can be applied in reducing the redundancy in the stock value information and leads to the reduction of highly correlated information to a reduced number of uncorrelated principal components, which normally account for several variances within the highly correlated information like in the stock market (Zhong & Enke, 2019).

DBSCAN

Machine learning can be used in numerous trading approaches. The density-based spatial clustering of application with noise, DBSCAN, could be applied to cluster the stocks and exclude the stocks that are not a fit for the group (Close & Kashef, 2020). The two applicable parameters for the method are the minPoints, which is the minimum number of points in the formation of a dense region, and eps, which is how close the points are supposed to be to each other to be part of the same cluster (Jeong & Park, 2019). The algorithm creates the clusters from the set of points that traders feed it from such parameters. Hence, the points within the low density are considered outliers.

Association rules learning

The association rule learning is the rule-based unsupervised machine learning strategy used in finding the interesting correlations, ‘IF THEN’ statements, within large datasets between variables (Sarker, 2021). The algorithm is based on various rules pf finding the interesting relations between various variables within the database. Some of the rules used include AIS and SETM, Apriori, Equivalence Class Clustering and bottom-up Lattice Traversal, and FP-Growth (Gayathri & Arunodhaya, 2021).

AIS algorithm

The AIS was the initial algorithmic rule developed mainly to produce all the large itemsets within the transaction database (Gireesha & Obulesu, 2017). The algorithm is essentially focused on the improvement of the datasets with the required functionality to process the decision support queries. Although the algorithm was created targeting the discovery of qualitative rule, the method to merely a single item within the consequent. The implication is that the association rule is mainly in the form of XÞI_j | a. Further, the algorithm creates several passes over the whole dataset, in which, in every pass, it scans all the transactions. Also, it is perceived that if the itemset is missing in the entire dataset, then it could never become a candidate for assessment of the huge itemsets within the succeeding pass. Thus, the items are stored in a lexicographic manner (Wankhede & Kamble, 2019). The traders use the AIS algorithm to establish any association between stock and their values, essential in forecasting. The technique needs more space and time to scan, thus cannot be used by traders to support quick decisions. Also, investors can use the technique to determine if there is an association between sectors within the clients’ buying behavior.

SETM algorithm

The algorithmic rule posits that every candidate itemset is produced on-the-fly mainly through database scanning and counting after the pass. Novel candidate itemsets are created similar to AIS, though the transaction identifier of the creating transaction is stored with the candidate itemset within the sequential structure. Hence, it separates candidate production procedures from counting (Mukherjee & Gupta, 2018). After every pass, the support count of every itemset is determined through the aggregation of the sequential structure. The technique has the same drawback as AIS, though for each candidate itemset, there are several entries as its support value (Al-Shamiri, 2021).

Apriori algorithm

The algorithm applies the concept that any subset of the larger itemset is expected to be a huge itemset. Besides, it presumes that items in the itemset are stored in a lexicographic manner. The basic distinction between this technique and that proposed by the SETM and AIS algorithms are the methods applied in the candidate itemset generation and the itemsets selection for counting (Panjaitan et al., 2019). Additionally, the technique generates the candidate itemset by fusing the huge itemsets of the earlier pass and removing all the subsets that are reduced within the past pass without deeming the dataset transaction. Thus, resulting in the reduced number of large candidates itemset. This forecasting method has two main challenges attributed to its intricate candidate generation procedure, requiring substantial space, memory, and time. Also, it necessitates several scans over the database to create the candidate itemset (Cheng et al., 2021). Therefore, the method can only be used for long-term decision-making and with experienced traders to assess certain associations in the stock market.

ECLAT Algorithm

The Equivalence Class Clustering and bottom-up Lattice Traversal, ECLAT, the algorithm is considered a more effective and scalable form of the Apriori algorithm (Sharma & Ganpati, 2021). Whereas the Apriori algorithm operates horizontally, imitating the Breadth-First Search of the graph, the ECLAT technique operates vertically, similar to the Depth-First Search of the graph, thus making it a faster technique compared to Apriori. The algorithm applies the basic concept of using the tidset intersection to calculate the support of the candidate itemset, hence averting the creation of subsets, which are nonexistent within the prefix tree (Devika, Koushik & Subramaniyaswamy, 2018). Therefore, it is considered the data mining technique initially created for market basket analysis (Lawrence, Mulyawan & Perdana, n/d). Hence the traders can use this algorithm in discovering the regularities in the client buying behavior in particular market segmentation. Essentially, the method identifies the set of services and products that are regularly purchased together, and once established, these sets of associated products can be utilized to enhance the firm productivity and could forecast the products that should be bundled or used in suggesting other products to clients (Das & Behera, 2017).

Frequency Pattern-Growth Algorithm

This technique is essentially the advancement of the apriori algorithm, which implies that the FP-growth rectifies the challenges associated with apriori techniques. Thus, it can be utilized in the assessment of the dataset that frequently appears within the dataset (Siahaan et al., 2017). The algorithm is designed to work on the dataset entailing transactions like the clients’ history of purchase in a specific market or organization. The bought item is considered ‘frequent.’ Hence, similar frequent would share the same branch of a tree, and when perceived to differ, the nodes will spit them (Anggraeni et al., 2019). Hence, the nodes identify the single items from the itemset, and the branch, path, represents the number of occurrences. Hence, the traders apply the algorithm in the swift discovery of the required frequency. Internally, the FP-growth algorithm does not necessitate the generation of the candidate since it applies the FP-tree data structure, making the technique efficient with large datasets (Ugwu & Udanor, 2021). The algorithm is divided into three sections, conditional pattern base, frequent itemset, and conditional FP-tress.

In summary, Artificial intelligence, especially machine learning, has advanced swiftly within the recent decades as aligned to its application in data analysis and computing, enabling the application to function intelligently. Further, machine learning often offers systems the capability of learning and improving from experience automatically without requiring to be programmed. An effective model of machine learning mainly relies on the information and the learning algorithms’ performance. Therefore, the complex algorithms undergo training by the gathered real-time information and knowledge linked to the prediction of stock value before the traders can use the system in making intelligent decisions regarding their investments. Besides, different machine learning methods, including supervised, unsupervised, and reinforced learning, play a critical role in ensuring accurate prediction and can be used with traders who possess an array of experience. Hence, choosing the suitable learning algorithm, which is perceived as appropriate for the stock value forecast, is often challenging because the purpose of distinct algorithms is extremely different. The results of the various algorithms within a similar group could differ reliant on the parameters used. Hence, it is critical for the traders to have insight on the principles of different machine learning algorithms and their applicability in order to successfully use them in predicting the stock values within a particular market.

References

Al-Shamiri, A. Y. R. (2021). Artificial intelligence and pattern recognition using data mining algorithms. International Journal of Computer Science & Network Security, 21(7), 221-232.

Anggraeni, S., Iha, M. A., Erawati, W., & Khairunnas, S. (2019). The Analysis of Sales by Using Apriori and FP-Growth at PT. Panca Putra Solusindo. REMIK: Riset dan E-Jurnal Manajemen Informatika Komputer, 3(2), 41-46.

Bini, B. S., & Mathew, T. (2016). Clustering and regression techniques for stock prediction. Procedia Technology, 24, 1248-1255.

Chen, Z., & Liu, W. (2020). An efficient parameter adaptive support vector regression using K-means clustering and chaotic slime mould algorithm. IEEE Access, 8, 156851-156862.

Cheng, K. C., Huang, M. J., Fu, C. K., Wang, K. H., Wang, H. M., & Lin, L. H. (2021). Establishing a multiple-criteria decision-making model for stock investment decisions using data mining techniques. Sustainability, 13(6), 3100.

Cheng, W. (2017). Statistical data mining for Sina Weibo, a Chinese micro-blog: Sentiment modelling and randomness reduction for topic modelling (Doctoral dissertation, London School of Economics and Political Science (LSE)).

Close, L., & Kashef, R. (2020). Combining artificial immune system and clustering analysis: A stock market anomaly detection model. Journal of Intelligent Learning Systems and Applications, 12(04), 83.

Das, K., & Behera, R. N. (2017). A survey on machine learning: concept, algorithms and applications. International Journal of Innovative Research in Computer and Communication Engineering, 5(2), 1301-1309.

Devika, R., Koushik, C., & Subramaniyaswamy, V. (2018). An Event Detection on Twitter using ECLAT (Equivalence Class Transformation) algorithm with TRCM (Transaction based Rule Change Mining). International Journal of Pure and Applied Mathematics, 119(12e), 13347-13356.

Fabregas, A. C., Gerardo, B. D., & Tanguilig III, B. T. (2017). Enhanced initial centroids for k-means algorithm. International Journal of Information Technology and Computer Science, 1, 26-33.

Fithri, F. A., & Wardhana, S. (2021). The CLUSTER ANALYSIS OF SALES TRANSACTION DATA USING K-MEANS CLUSTERING AT TOKO USAHA MANDIRI. Pilar Nusa Mandiri: Journal of Computing and Information System, 17(2), 113-118.

Gayathri, K., & Arunodhaya, R. (2021). Customer Segmentation and Personalized Marketing Using K-Means and APRIORI Algorithm.

Gireesha, O., & Obulesu, O. (2017). TKAR: Efficient Mining of Top-k Association Rules on Real—Life Datasets. In Proceedings of the 5th International Conference on Frontiers in Intelligent Computing: Theory and Applications (pp. 45-54). Springer, Singapore.

Huang, D., Wang, C. D., Wu, J. S., Lai, J. H., & Kwoh, C. K. (2019). Ultra-scalable spectral clustering and ensemble clustering. IEEE Transactions on Knowledge and Data Engineering, 32(6), 1212-1226.

Jeong, Y. S., & Park, J. H. (2019). Learning algorithms in AI system and services. Journal of Information Processing Systems, 15(5), 1029-1035.

Kerenidis, I., & Landman, J. (2021). Quantum spectral clustering. Physical Review A, 103(4), 042415.

Lawrence, E., Mulyawan, B., & Perdana, N. J. (n/d). MARKET BASKET ANALYST BASED ON WEBSITE USING ECLAT ALGORITHM (CASE STUDY POLA PHARMACY). Jurnal Ilmu Komputer dan Sistem Informasi, 8(2), 205-208.

Lu, J., & Zhu, Q. (2017). An effective algorithm based on density clustering framework. Ieee Access, 5, 4991-5000.

Mukherjee, S., & Gupta, G. D. P. (2018). Integrating Application with Algorithms of Association Rule used in Descriptive Data Modelling, through which Data Mining can be Implemented for Future Prediction. International Journal of Applied Engineering Research, 13(17), 13272-13281.

Pahwa, N., Khalfay, N., Soni, V., & Vora, D. (2017). Stock prediction using machine learning a review paper. International Journal of Computer Applications, 163(5), 36-43.

Panjaitan, S., Amin, M., Lindawati, S., Watrianthos, R., Sihotang, H. T., & Sinaga, B. (2019, August). Implementation of apriori algorithm for analysis of consumer purchase patterns. In Journal of Physics: Conference Series (Vol. 1255, No. 1, p. 012057). IOP Publishing.

Prasdika, P., & Sugiantoro, B. (2018). A review paper on big data and data mining concepts and techniques. IJID (International Journal on Informatics for Development), 7(1), 36-38.

Roux, M. (2018). A comparative study of divisive and agglomerative hierarchical clustering algorithms. Journal of Classification, 35(2), 345-366.

Sakhare, N. N., & Imambi, S. S. (2019). Performance analysis of regression based machine learning techniques for prediction of stock market movement. International Journal of Recent Technology and Engineering, 7(6), 655-662.

Sarker, I. H. (2021). Machine learning: Algorithms, real-world applications and research directions. SN Computer Science, 2(3), 1-21.

Shah, D., Isah, H., & Zulkernine, F. (2019). Stock market analysis: A review and taxonomy of prediction techniques. International Journal of Financial Studies, 7(2), 26.

Sharma, A., & Ganpati, A. (2021). Association rule mining algorithms: A Comparative review. International Research Journal of Engineering and Technology, 8(11).

Shetty, P., & Singh, S. (2021). Hierarchical Clustering: A Survey. International Journal of Applied Research, 7(4), 178-181.

Siahaan, A. P. U., Mesran, M., Lubis, A. H., & Ikhwan, A. (2017). Association Rules Analysis on FP-Growth Method in Predicting Sales.

Sinaga, K. P., & Yang, M. S. (2020). Unsupervised K-means clustering algorithm. IEEE access, 8, 80716-80727.

Singh, J., & Khushi, M. (2021). Feature Learning for Stock Price Prediction Shows a Significant Role of Analyst Rating. Applied System Innovation, 4(1), 17.

Sokol, O. (2019). Hierarchical clustering of products using market-basket data. In International Conference on Advances in Business and Law (ICABL) (Vol. 3, No. 1, pp. 88-93).

Swarndeep Saket, J., & Pandya, S. (2016). An overview of partitioning algorithms in clustering techniques. International Journal of Advanced Research in Computer Engineering & Technology (IJARCET), 5(6), 1943-1946.

Tremblay, N., & Loukas, A. (2020). Approximating spectral clustering via sampling: a review. Sampling Techniques for Supervised or Unsupervised Tasks, 129-183.

Ugwu, N. V., & Udanor, C. N. (2021). Achieving effective customer relationship using frequent pattern-growth algorithm association rule learning technique. Nigerian Journal of Technology, 40(2), 329-339.

Wankhede, P., & Kamble, V. (2019). Efficient Algorithm for Frequent Item Set Generation in Big Data.

Wen, Y., Lin, P., & Nie, X. (2020, March). Research of stock price prediction based on PCA-LSTM model. In IOP Conference Series: Materials Science and Engineering (Vol. 790, No. 1, p. 012109). IOP Publishing.

Zhong, X., & Enke, D. (2019). Predicting the daily return direction of the stock market using hybrid machine learning algorithms. Financial Innovation, 5(1), 1-20.

Zhu, X., Zhang, S., He, W., Hu, R., Lei, C., & Zhu, P. (2018). One-step multi-view spectral clustering. IEEE Transactions on Knowledge and Data Engineering, 31(10), 2022-2034.

Zhu, Z., & Liu, N. (2021). Early warning of financial risk based on the k-means clustering algorithm. Complexity, 2021.